Presentation is loading. Please wait.

Presentation is loading. Please wait.

Data-Parallel Programming Model Basic uniform operations across lattice: C(x) = A(x)*B(x) Distribute problem grid across a machine grid Want API to hide.

Similar presentations


Presentation on theme: "Data-Parallel Programming Model Basic uniform operations across lattice: C(x) = A(x)*B(x) Distribute problem grid across a machine grid Want API to hide."— Presentation transcript:

1 Data-Parallel Programming Model Basic uniform operations across lattice: C(x) = A(x)*B(x) Distribute problem grid across a machine grid Want API to hide subgrid layout and communications (bundling subgrid faces) between nodes Data-parallel – basic idea something like Fortran 90 with more complex types Implement via a programming interface (API), e.g. a C interface Data layout over processors

2 QCD Data Types Fields have various types (indices): Lattice Fields or scalar (same on all sites) Index type (at a site) Gauge fields:Product(Matrix(Nc),Scalar) Fermions:Product(Vector(Nc),Vector(Ns)) Scalars:Scalar Propagators:Product(Matrix(Nc),Matrix(Ns)) Support compatible operations on types: Matrix(color)*Matrix(spin)*Vector(color,spin)

3 QCD-API Dirac Operators, CG Routines etc. C, C++, etc. (Organized by high level codes etc.) Level 3 Data Parallel QCD Lattice Operations (overlapping Algebra and Messaging) A = SHIFT(B, mu) * C; Global sums, etc Level 2 Lattice Wide Lin Alg (No Communication) e.g. A = B * C Lattice Wide Data Movement (Blocking Communication) e.g Atemp = SHIFT(A, mu) Level 1 QLA: Linear Algebra API SU(3), gamma algebra etc. QMP: Message Passing API Communications I/O (XML) Serialization/ Marshalling

4 Level 2 Functionality Object types/declarations: LatticeGaugeF, ReadF, ComplexF, … Unary operations: operate on one source and return (or into) a target void Shift(LatticeGaugeF target, LatticeGaugeF source, int sign, int direction, Subset s); void Copy(LatticeFermionF dest, LatticeFermionF source, Subset s); void Trace(LatticeComplexF dest, LatticeGaugeF source, Subset s); Binary operations: operate on two sources and return into a target void Multiply(LatticePropagatorF dest, LatticePropagatorF src1, LatticeGaugeF src2, Subset s); void Gt(LatticeBooleanF dest, LatticeRealF src1, LatticeRealF src2, Subset s); Broadcasts: broadcast throughout lattice void Fill(LatticeGaugeF dest, RealF val, Subset s); Reductions: reduce through the lattice void Norm2(RealD dest, LatticeFermionF source, Subset s);

5 C Interface for Level 2 Binary: Multiplication has many varieties void QDP_F_T3_eqop_T1op1_mult_T2op2 (restrict Type3 *r, Type1 *a, Type2 *b, const Subset s) Multiply operations where T1, T2, T3 are shortened type names for the types Type1, Type2 and Type3 LatticeGaugeF, LatticeDiracFermionF, LatticeHalfFermionF, LatticePropagatorF, ComplexF, LatticeComplexF, RealF, LatticeRealF, … and eqop,op1,op2 considered as a whole imply operations like eq,,r = a*beqm,,r = -a*b eq,,Ar = a*conj(b)eqm,,Ar = -a*conj(b) eq,A,r = conj(a)*beqm,A,r = -conj(a)*b eq,A,Ar = conj(a)*conj(b)eqm,A,Ar = -conj(a)*conj(b) peq,,r = r + a*bmeq,,r = r – a*b peq,,Ar = r + a*conj(b)meq,,Ar = r – a*conj(b) peq,A,r = r + conj(a)*bmeq,A,r = r – conj(a)*b peq,A,Ar = r + conj(a)*conj(b)meq,A,Ar = r - conj(a)*conj(b)

6 QDP++ Why C++ ? –Many OOP’s benefits: Simplify naming with operator overloading Strong type checking – hard to call wrong routine Can create expressions! Closures straightforward - return types easily Can straightforwardly automate memory allocation/destruction Good support tools (e.g., doxygen – documentation generator) –Maintenance: Generate large numbers of variant routines via templates Contrary to popular opinion, there are compilers that meet standards. Minimum level is GNU g++. Expect to use full 3.0 standards Simple to replace crucial sections (instantiations) with optimized code Initial tests – g++ generating decent assembly! –User acceptance: Language in common use

7 Flavors Operation syntax comes in two variants: Global functional form (optimization straightforward): void Multiply_replace(Type3& dest, const Type1& src1, const Type2& src2); Operator form (optimizations require delayed evaluation): Type3 operator*(const Type1& src1, const Type2& src2); void operator=(Type3& dest, const Type3& src1); Were the types Type1, Type2 and Type3 are the names again LatticeGaugeF, LatticeDiracFermionF, … Declarations: Objects declared and allocated across the lattice (not initialized) LatticeGaugeF a, b;

8 Templates for Operations Without templates, multiple overloaded declarations required: void Multiply_op3(LatticeComplexF& dest, const RealF& src1, const LatticeComplexF& src2); void Multiply_op3(LatticeDiracFermionF& dest, const LatticeGaugeF& src1, const LatticeDiracFermionF& src2); Can merge similar operations with templates: template void Multiply_op3(T3& dest, const T1& src1, const T2& src2); Optimizations easily provided via specializations: template void Multiply_op3(LatticeDiracFermionF& dest, const LatticeGaugeF& src1, const LatticeDiracFermionF& src2) {/* Call your favorite QLA or QDP routine here */}

9 Templates for Declarations Expect to have data driven (and not method driven) class types: class LatticeDiracFermionF { private: DiracFermionF sites[layout.Volume()]; }; With templates: template class Lattice { private: T sites[layout.Volume()]; }; typedef Lattice LatticeDiracFermionF; Type composition: The fiber (site) types can themselves be made from templates (implementation choice)

10 Implementation Example Use a container class to hold info on object Easily separate functional and operator syntatic forms Closures implemented at this container level Underlying object class does real work No additional overhead – use inlining QLA – optimized linear algebra routines QMP – message passing / communications QDP object syntax Operator syntax Functional syntax // Functional form inline void Multiply_replace(QDP & dest, const QDP & src1, const QDP & src2) { dest.obj().mult_rep(src1.obj(), src2.obj()); } // Operator form without closure optimizations inline QDP operator*(const QDP & src1, const QDP & src2) {QDP tmp; tmp.obj().mult_rep(src1.obj(), src2.obj()); return tmp;}

11 Example // Declarations and construction – no default initialization LatticeGaugeF u, tmp0, tmp1; // Initialization Gaussian(u); Zero(tmp0); tmp1 = 1.0; // Two equivalent examples Multiply_replace(tmp0, u, tmp1); tmp0 = u * tmp1; // Three equivalent examples Multiply_conj2_ shift2_ replace(tmp1, u, tmp0, FORWARD, mu); Multiply_replace(tmp1, u, Conj(Shift(tmp0,FORWARD,mu))); tmp1 = u * Conj(Shift(tmp0,FORWARD,mu)); { // Change default subset as a side-effect of a context declaration Context foo(Even_subset); tmp1 = u * Conj(Shift(tmp0,FORWARD,mu)); // Only on even subset // Context popped on destruction }

12 SZIN SZIN and the Art of Software Maintenance –M4 preprocessed object-oriented data-parallel programming interface –Base code is C –High level code implemented over architectural dependent level 2 layer –Supports scalar nodes, multi-process SMP, multi-threaded SMP, vector, CM-2, CM-5, clusters of SMP’s –The Level 2 routines are generated at M4 time into a repository –Supports overlapping communications/computations –No exposed Level 1 layer –Can/does use (transparent to user) any external Level 1 or 2 routines –Optimized Level 3 routines (Dirac operator) available –Code publicly available by tar and CVS with many contributors – no instances of separate code camps Data parallel interface –SZIN is really the data-parallel interface and not the high level code –Some projects have only used the low level code and library –Could easily produce a standalone C implementation of the Level 2 API

13 Problems with SZIN Why change current system? Cons: –Maintenance: M4 has no scope. Preprocessor run over a file - no knowledge of C syntax or scope – bizarre errors can result M4 has no programming support for holding type info. Must invent it Interface problem – awkward to hook onto level 1 like optimizations –Extensibility: Problems with C – awkward to write generic algorithms, e.g. CG for different Dirac operators (like Clover requiring sigma.F but not Wilson), or algorithms on lattices or just single sites. No real language support for closures C very limited on type checking No automated memory allocation/free-ing. No simple way to write expressions. –User acceptance: M4 not familiar (and limited) to most users

14 Current Development Efforts QDP++ development –Level-2 C/C++ API not yet finalized by SciDAC – documentation underway –Currently testing the basic object model code needed for the single node version –Shortly start the parallel version – uses object model code –Will incorporate QLA library routines once agreed upon by SciDAC Stand-alone demonstration/prototyping suite –Some test routines and drivers for physics routines live in the QDP++ test directory –David/Richard/Andrew/others writing some simple stand-alone routines (like pure gauge) as a prototyping effort –Expect to produce a user’s guide SZIN port –Porting existing SZIN code with merging of MIT and Maryland code bases –The M4 architectural support part is completely removed and only uses QDP –Will become another high level code base using implementation dependent QDP

15 Hardware Acquisitions FY02 –NOW: 128 single processor P4, 2Ghz, Myrinet Expect ~ 130 Gflops –Late summer: probably 256 single processor P4, 2Ghz, 3-dimensional Gigabit-ethernet Expect  200 Gflops FY03 –Late summer: probably 256 single processor P4, 2.4Ghz, 3-dimensional Gigabit-ethernet Expect  240 Gflops


Download ppt "Data-Parallel Programming Model Basic uniform operations across lattice: C(x) = A(x)*B(x) Distribute problem grid across a machine grid Want API to hide."

Similar presentations


Ads by Google